Open Access

Sebastian Möller

• Based on earlier projects [Hei20, HM22], a CV-pipeline for segmentation and identification of single cows in images was created. The obtained results led to the idea of converting the given used server-based approach to end-devices like a recent iOS-based phone or tablet with its built-in machine learning accelerators to examine if such an application is feasible for real-life usage, especially regarding accuracy and performance. The further goal behind this project is to give farmers and veterinarians a tool to instantly identify single animals based on a photo taken from behind with the integrated camera of their end-device like e.g., an iPhone. This is useful because a cow is normally medicated seized in a milking facility where its ear tag is not easily accessible. In this use case a high identification accuracy >90% is mandatory to avoid treatment errors. This project's topic is the development of a native iOS application to identify single cows based on photos and evaluate the performance and quality of end-device inference with a recent iOS device. Therefore, the application makes use of Apple's Core ML Framework, especially of the 'Vision' part for working with image-based data. The used CV-models are partially translated from PyTorch and TensorFlow via Apple's coremltools and in comparison, a completely new identification model was created with Xcode's Create ML Application.